From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from NAM12-MW2-obe.outbound.protection.outlook.com (mail-mw2nam12on2089.outbound.protection.outlook.com [40.107.244.89]) by gabe.freedesktop.org (Postfix) with ESMTPS id 7623D10E3CA for ; Tue, 29 Aug 2023 17:18:52 +0000 (UTC) Message-ID: <55ec5964-f2f4-4e14-b3ff-b880c3dcb6ad@amd.com> Date: Tue, 29 Aug 2023 13:18:46 -0400 To: vitaly.prosyak@amd.com, igt-dev@lists.freedesktop.org References: <20230829044006.230513-1-vitaly.prosyak@amd.com> Content-Language: en-CA, en-US From: Luben Tuikov In-Reply-To: <20230829044006.230513-1-vitaly.prosyak@amd.com> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit MIME-Version: 1.0 Subject: Re: [igt-dev] [PATCH] tests/amdgpu: add sync object tests List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: Alex Deucher , Christian Koenig Errors-To: igt-dev-bounces@lists.freedesktop.org Sender: "igt-dev" List-ID: Hi Vitaly, Good work--thanks for working on this. Inlined: On 2023-08-29 00:40, vitaly.prosyak@amd.com wrote: > From: Vitaly Prosyak > > Using worker thread to wait on point and then signal point on other thread. > Another test uses a worker thread to signal point and wait on the main > thread using amdgpu_cs_syncobj_timeline_wait. > > The command consists of two chunks : > 1. AMDGPU_CHUNK_ID_IB uses GFX_COMPUTE_NOP or SDMA_NOP. > 2. The second chunk is AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT > or AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL which has the > point number . > > v2 : Kamil helped with podman script to fix the build > failure when drmlib version < 2.4.97 > > Cc: Luben Tuikov > Cc: Alex Deucher > Cc: Christian Koenig > Cc: Kamil Konieczny > Signed-off-by: Vitaly Prosyak > --- > tests/amdgpu/amd_syncobj.c | 262 +++++++++++++++++++++++++++++++++++++ > tests/amdgpu/meson.build | 5 + > 2 files changed, 267 insertions(+) > create mode 100644 tests/amdgpu/amd_syncobj.c > > diff --git a/tests/amdgpu/amd_syncobj.c b/tests/amdgpu/amd_syncobj.c > new file mode 100644 > index 000000000..e83b5c5ad > --- /dev/null > +++ b/tests/amdgpu/amd_syncobj.c > @@ -0,0 +1,262 @@ > +// SPDX-License-Identifier: MIT > + Should probably include the standard AMD copyright notice. /* Copyright (C) 2023 Advanced Micro Devices Inc. */ > +#include > +#include > +#include > + > +#include "igt.h" > +#include "lib/amdgpu/amd_PM4.h" > +#include "lib/amdgpu/amd_sdma.h" > +#include "lib/amdgpu/amd_memory.h" > + > +struct syncobj_point { > + amdgpu_device_handle device; > + uint32_t syncobj_handle; > + uint64_t point; > +}; > + > + > +static bool > +syncobj_timeline_enable(int fd) > +{ > + int r; > + bool ret = false; > + uint64_t cap = 0; > + > + r = drmGetCap(fd, DRM_CAP_SYNCOBJ_TIMELINE, &cap); > + if (r || cap == 0) > + return ret; > + ret = true; > + > + return ret; > +} In the function above, you don't need "ret"--the compiler optimizes it away as follows: static bool syncobj_timeline_enable(int fd) { uint64_t cap = 0; int r; r = drmGetCap(fd, DRM_CAP_SYNCOBJ_TIMELINE, &cap); return !(r || cap == 0); } Or you can use "!!!(r || cap == 0);" if the/a compiler complains. > + > +static void > +syncobj_command_submission_helper(amdgpu_device_handle device_handle, > + uint32_t syncobj_handle, bool wait_or_signal, uint64_t point) > +{ > + amdgpu_context_handle context_handle; > + amdgpu_bo_handle ib_result_handle; > + void *ib_result_cpu; > + uint64_t ib_result_mc_address; > + struct drm_amdgpu_cs_chunk chunks[2]; > + struct drm_amdgpu_cs_chunk_data chunk_data; > + struct drm_amdgpu_cs_chunk_syncobj syncobj_data; > + struct amdgpu_cs_fence fence_status; > + amdgpu_bo_list_handle bo_list; > + amdgpu_va_handle va_handle; > + uint32_t expired; > + int i, r; > + uint64_t seq_no; > + uint32_t *ptr; > + > + r = amdgpu_cs_ctx_create(device_handle, &context_handle); > + igt_assert_eq(r, 0); > + > + r = amdgpu_bo_alloc_and_map(device_handle, 4096, 4096, > + AMDGPU_GEM_DOMAIN_GTT, 0, > + &ib_result_handle, &ib_result_cpu, > + &ib_result_mc_address, &va_handle); > + igt_assert_eq(r, 0); > + > + r = amdgpu_get_bo_list(device_handle, ib_result_handle, NULL, &bo_list); > + igt_assert_eq(r, 0); > + > + ptr = ib_result_cpu; > + > + for (i = 0; i < 16; ++i) > + ptr[i] = wait_or_signal ? GFX_COMPUTE_NOP : SDMA_NOP; > + > + chunks[0].chunk_id = AMDGPU_CHUNK_ID_IB; > + chunks[0].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4; > + chunks[0].chunk_data = (uint64_t)(uintptr_t)&chunk_data; > + chunk_data.ib_data._pad = 0; > + chunk_data.ib_data.va_start = ib_result_mc_address; > + chunk_data.ib_data.ib_bytes = 16 * 4; > + chunk_data.ib_data.ip_type = wait_or_signal ? AMDGPU_HW_IP_GFX : AMDGPU_HW_IP_DMA; > + chunk_data.ib_data.ip_instance = 0; > + chunk_data.ib_data.ring = 0; > + chunk_data.ib_data.flags = 0; > + > + chunks[1].chunk_id = wait_or_signal ? > + AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT : > + AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL; > + chunks[1].length_dw = sizeof(struct drm_amdgpu_cs_chunk_syncobj) / 4; > + chunks[1].chunk_data = (uint64_t)(uintptr_t)&syncobj_data; > + syncobj_data.handle = syncobj_handle; > + syncobj_data.point = point; > + syncobj_data.flags = DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT; > + > + r = amdgpu_cs_submit_raw(device_handle, > + context_handle, > + bo_list, > + 2, > + chunks, > + &seq_no); > + igt_assert_eq(r, 0); > + > + memset(&fence_status, 0, sizeof(struct amdgpu_cs_fence)); > + fence_status.context = context_handle; > + fence_status.ip_type = wait_or_signal ? AMDGPU_HW_IP_GFX : AMDGPU_HW_IP_DMA; > + fence_status.ip_instance = 0; > + fence_status.ring = 0; > + fence_status.fence = seq_no; > + > + r = amdgpu_cs_query_fence_status(&fence_status, > + AMDGPU_TIMEOUT_INFINITE, 0, &expired); > + igt_assert_eq(r, 0); > + > + r = amdgpu_bo_list_destroy(bo_list); > + igt_assert_eq(r, 0); > + > + amdgpu_bo_unmap_and_free(ib_result_handle, va_handle, > + ib_result_mc_address, 4096); > + > + r = amdgpu_cs_ctx_free(context_handle); > + igt_assert_eq(r, 0); > +} > + > +static void * > +syncobj_wait(void *data) > +{ > + struct syncobj_point *sp = (struct syncobj_point *)data; > + > + syncobj_command_submission_helper(sp->device, sp->syncobj_handle, true, > + sp->point); > + > + return (void *)0; > +} > + > +static void * > +syncobj_signal(void *data) > +{ > + struct syncobj_point *sp = (struct syncobj_point *)data; > + > + syncobj_command_submission_helper(sp->device, sp->syncobj_handle, false, > + sp->point); > + > + return (void *)0; > +} > + > +static void > +amdgpu_syncobj_timeline(amdgpu_device_handle device_handle) > +{ > + static pthread_t wait_thread; > + static pthread_t signal_thread; > + static pthread_t c_thread; > + struct syncobj_point sp1, sp2, sp3; > + uint32_t syncobj_handle; > + uint64_t payload; > + uint64_t wait_point, signal_point; > + uint64_t timeout; > + struct timespec tp; > + int r, sync_fd; > + void *tmp, *tmp2; > + > + r = amdgpu_cs_create_syncobj2(device_handle, 0, &syncobj_handle); > + igt_assert_eq(r, 0); > + > + // wait on point 5 > + sp1.syncobj_handle = syncobj_handle; > + sp1.device = device_handle; > + sp1.point = 5; > + r = pthread_create(&wait_thread, NULL, syncobj_wait, &sp1); > + igt_assert_eq(r, 0); > + > + // signal on point 10 > + sp2.syncobj_handle = syncobj_handle; > + sp2.device = device_handle; > + sp2.point = 10; > + r = pthread_create(&signal_thread, NULL, syncobj_signal, &sp2); > + igt_assert_eq(r, 0); > + > + r = pthread_join(signal_thread, &tmp); > + igt_assert_eq(r, 0); > + > + r = pthread_join(wait_thread, &tmp2); > + igt_assert_eq(r, 0); > + > + //query timeline payload > + r = amdgpu_cs_syncobj_query(device_handle, &syncobj_handle, > + &payload, 1); > + igt_assert_eq(r, 0); > + igt_assert_eq(payload, 10); > + > + //signal on point 16 > + sp3.syncobj_handle = syncobj_handle; > + sp3.device = device_handle; > + sp3.point = 16; > + r = pthread_create(&c_thread, NULL, syncobj_signal, &sp3); > + igt_assert_eq(r, 0); > + > + //CPU wait on point 16 > + wait_point = 16; > + timeout = 0; > + clock_gettime(CLOCK_MONOTONIC, &tp); > + timeout = tp.tv_sec * 1000000000ULL + tp.tv_nsec; > + timeout += 10000000000; //10s > + r = amdgpu_cs_syncobj_timeline_wait(device_handle, &syncobj_handle, > + &wait_point, 1, timeout, > + DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL | > + DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT, > + NULL); > + > + igt_assert_eq(r, 0); > + r = pthread_join(c_thread, &tmp); > + igt_assert_eq(r, 0); > + > + // export point 16 and import to point 18 > + r = amdgpu_cs_syncobj_export_sync_file2(device_handle, syncobj_handle, > + 16, > + DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT, > + &sync_fd); > + igt_assert_eq(r, 0); > + r = amdgpu_cs_syncobj_import_sync_file2(device_handle, syncobj_handle, > + 18, sync_fd); > + igt_assert_eq(r, 0); > + r = amdgpu_cs_syncobj_query(device_handle, &syncobj_handle, > + &payload, 1); > + igt_assert_eq(r, 0); > + igt_assert_eq(payload, 18); > + > + // CPU signal on point 20 > + signal_point = 20; > + r = amdgpu_cs_syncobj_timeline_signal(device_handle, &syncobj_handle, > + &signal_point, 1); > + igt_assert_eq(r, 0); > + r = amdgpu_cs_syncobj_query(device_handle, &syncobj_handle, > + &payload, 1); > + igt_assert_eq(r, 0); > + igt_assert_eq(payload, 20); > + > + r = amdgpu_cs_destroy_syncobj(device_handle, syncobj_handle); > + igt_assert_eq(r, 0); > + > +} Very nice. :-) One question: Is it possible we hang indefinitely in pthread_create()-->syncobj_signal/wait()-->amdgpu_cs_query_fence_status()? If yes, is there a way to mitigate this and timeout the test, if run from a nightly script for instance? > + > +igt_main > +{ > + amdgpu_device_handle device; > + int fd = -1; > + > + igt_fixture { > + uint32_t major, minor; > + int err; > + > + fd = drm_open_driver(DRIVER_AMDGPU); > + err = amdgpu_device_initialize(fd, &major, &minor, &device); > + igt_require(err == 0); > + igt_require(syncobj_timeline_enable(fd)); > + igt_info("Initialized amdgpu, driver version %d.%d\n", > + major, minor); > + > + } > + > + igt_subtest("amdgpu_syncobj_timeline") > + amdgpu_syncobj_timeline(device); > + > + igt_fixture { > + amdgpu_device_deinitialize(device); > + close(fd); > + } > +} > diff --git a/tests/amdgpu/meson.build b/tests/amdgpu/meson.build > index 6032a38e8..1520c34bd 100644 > --- a/tests/amdgpu/meson.build > +++ b/tests/amdgpu/meson.build > @@ -37,6 +37,11 @@ if libdrm_amdgpu.found() > 'amd_mall', > 'amd_odm', > ] > + if libdrm_amdgpu.version().version_compare('> 2.4.97') > + amdgpu_progs +=[ 'amd_syncobj', ] > + else > + warning('libdrm <= 2.4.97 found, no amd_syncobj test') I'd probably print: 'libdrm <= 2.4.97 found, amd_syncobj test not applicable' to clarify that for libdrm versions less than or equal to the one indicated, the test while available is not applicable. > + endif > amdgpu_deps += libdrm_amdgpu > endif > With these changes applied, this patch is, Reviewed-by: Luben Tuikov -- Regards, Luben